{
  "id": 77605,
  "title": "New Paper on Dolphin Identification using Triplet Loss",
  "url": "/competitions/humpback-whale-identification/discussion/77605",
  "author_name": "",
  "post_date": "2019-01-14T19:41:56.217501200Z",
  "votes": 18,
  "comment_count": 2,
  "views": 0,
  "content": "<p>I was reading through arxiv.org and saw this recent paper that I thought might prove to be a good overview of applying triplet loss to a similar dataset.  While i don't see anything incredibly unique about their approach, it may be helpful to check it over to compare methodologies.</p>\n\n<p><a href=\"https://arxiv.org/abs/1901.03662\">Individual common dolphin identification via metric embedding learning</a></p>\n\n<p>-Paul</p>",
  "messages": [
    {
      "id": "455892",
      "postDate": "01/14/2019 19:41:56",
      "content": "<p>I was reading through arxiv.org and saw this recent paper that I thought might prove to be a good overview of applying triplet loss to a similar dataset.  While i don't see anything incredibly unique about their approach, it may be helpful to check it over to compare methodologies.</p>\n\n<p><a href=\"https://arxiv.org/abs/1901.03662\">Individual common dolphin identification via metric embedding learning</a></p>\n\n<p>-Paul</p>",
      "rawMarkdown": "I was reading through arxiv.org and saw this recent paper that I thought might prove to be a good overview of applying triplet loss to a similar dataset.  While i don't see anything incredibly unique about their approach, it may be helpful to check it over to compare methodologies.\n\n[Individual common dolphin identification via metric embedding learning][1]\n\n\n  [1]: https://arxiv.org/abs/1901.03662\n\n-Paul",
      "votes": null
    },
    {
      "id": "456736",
      "postDate": "01/16/2019 12:22:12",
      "content": "<p>Interesting find! If I understand this loss function correctly, you train it by picking a sample, then you compare it to a match, and then you compare it to a mismatch. Those are the triplets. Overall, you're trying to lean towards matches and away from mismatches. I wonder if this makes it Three Shot Learning? </p>\n\n<p>Here's a brief explanation of <a href=\"https://machinelearning.wtf/terms/triplet-loss/\">Triplet Loss</a>.</p>",
      "rawMarkdown": "Interesting find! If I understand this loss function correctly, you train it by picking a sample, then you compare it to a match, and then you compare it to a mismatch. Those are the triplets. Overall, you're trying to lean towards matches and away from mismatches. I wonder if this makes it Three Shot Learning? \n\nHere's a brief explanation of [Triplet Loss](https://machinelearning.wtf/terms/triplet-loss/).",
      "votes": null
    },
    {
      "id": "460228",
      "postDate": "01/23/2019 08:25:54",
      "content": "<p>As far as I'm concerned, N-shot learning refers rather to the amount of training data available for a class than to the number of examples necessary for loss calculation</p>\n\n<p>Btw, here is another <a href=\"https://omoindrot.github.io/triplet-loss\">link</a> that might be useful.</p>",
      "rawMarkdown": "As far as I'm concerned, N-shot learning refers rather to the amount of training data available for a class than to the number of examples necessary for loss calculation\n\nBtw, here is another [link][1] that might be useful.\n\n\n  [1]: https://omoindrot.github.io/triplet-loss",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 456736,
      "author_name": "devilears",
      "author_url": "",
      "post_date": "01/16/2019 12:22:12",
      "content": "<p>Interesting find! If I understand this loss function correctly, you train it by picking a sample, then you compare it to a match, and then you compare it to a mismatch. Those are the triplets. Overall, you're trying to lean towards matches and away from mismatches. I wonder if this makes it Three Shot Learning? </p>\n\n<p>Here's a brief explanation of <a href=\"https://machinelearning.wtf/terms/triplet-loss/\">Triplet Loss</a>.</p>",
      "votes": null,
      "replies": [
        {
          "id": 460228,
          "author_name": "devvindan",
          "author_url": "",
          "post_date": "01/23/2019 08:25:54",
          "content": "<p>As far as I'm concerned, N-shot learning refers rather to the amount of training data available for a class than to the number of examples necessary for loss calculation</p>\n\n<p>Btw, here is another <a href=\"https://omoindrot.github.io/triplet-loss\">link</a> that might be useful.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "455892": "I was reading through arxiv.org and saw this recent paper that I thought might prove to be a good overview of applying triplet loss to a similar dataset.  While i don't see anything incredibly unique about their approach, it may be helpful to check it over to compare methodologies.\n\n[Individual common dolphin identification via metric embedding learning][1]\n\n\n  [1]: https://arxiv.org/abs/1901.03662\n\n-Paul",
    "456736": "Interesting find! If I understand this loss function correctly, you train it by picking a sample, then you compare it to a match, and then you compare it to a mismatch. Those are the triplets. Overall, you're trying to lean towards matches and away from mismatches. I wonder if this makes it Three Shot Learning? \n\nHere's a brief explanation of [Triplet Loss](https://machinelearning.wtf/terms/triplet-loss/).",
    "460228": "As far as I'm concerned, N-shot learning refers rather to the amount of training data available for a class than to the number of examples necessary for loss calculation\n\nBtw, here is another [link][1] that might be useful.\n\n\n  [1]: https://omoindrot.github.io/triplet-loss"
  },
  "source": "meta"
}